Otherwise they're oriented vertically like this:
(In BC7 terminology, ETC1 supports 1 or 2 subsets per block, with 1 or 2 partitions.)
Anyhow, a high quality encoder typically tries both subblock orientations and chooses the one which results in the lowest overall error. Interestingly, it's possible to predict with a reasonable degree of certainty which subblock orientation is best. The advantage to this method is obvious: higher encoding throughput, with (hopefully) only a small loss in quality.
The method used by etc2comp computes each possible subblock's average color, then it sums the squared "gray distance" from each pixel to each subblock's average color. (To compute gray distance from color A vs. B, it projects A onto the grayscale line going through B, then it computes the squared RGB distance from A to this projected point.) etc2comp then chooses the flip orientation with the lowest overall gray distance to try first.
Success rates on a handful of images, with RGB avg. PSNR and SSIM stats (using accurate vs. estimated flips):
kodim01: 68% (35.904 vs. 35.607, .975236 vs. .973744)
kodim03: 65% (39.949 vs. 38.801, .964774 vs. .963467)
kodim18: 69% (35.917 vs. 35.661, .965767 vs. .964385)
kodim20: 63% (38.849 vs. 38.589, .975558 vs. .974669)
ETC1 block flip bit visualizations on kodim03:
basislib best flips:
Flip estimates, using etc2comp's gray line estimate algorithm:
etc2comp's flips (in ETC1 mode):
etcpak's flips (in ETC1 mode):